tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Hierarchical STFT based Transformer for Causality discovery

Authors :

Sahar Semsarha¹ Mohammad bagher Shamsolahi²

1- sharif university of technology 2- sharif university of techonolgy

Keywords :

EEG،causal discovery،Transformers،connectivity،time-frequency analysis

Abstract :

Abstract— electroencephalogram (EEG) signal analysis is crucial for understanding brain dynamics and connectivity. Traditional approaches such as Granger causality, Partial Directed Coherence (PDC), and Directed Transfer Function (DTF) rely on linear autoregressive assumptions and often fail to capture nonlinear dependencies. At the same time, deep learning models including CNNs, RNNs, and Transformers have achieved strong results in EEG decoding tasks, yet these methods generally focus on correlation rather than causation. To address these limitations, we propose a Hierarchical Causal-STFT Transformer (H-STFT-T), a novel framework that integrates causal short-time Fourier transform (STFT) representations with a multi-level hierarchical Transformer architecture. By enforcing causality in the spectral domain and incorporating intra-patch, inter-patch, and inter-channel attention modules, our model prevents leakage, learns temporal delays (lags), and generates directed connectivity graphs. We evaluate H-STFT-T on both synthetic datasets and real EEG benchmarks. Experimental results demonstrate that our method achieves superior accuracy in recovering ground-truth causal links and lag structures, outperforming classical approaches (Granger, PDC, DTF), deep non-causal baselines (GCN, GAT, Transformer), and causal discovery methods (PCMCI, Transfer Entropy, LiNGAM).